Smoothie Qwen

Smoothie Qwen is a lightweight adjustment tool that smooths token probabilities in Qwen and similar models, enhancing balanced multilingual generation capabilities. For more details, please refer to https://github.com/dnotitia/smoothie-qwen.

Configuration

• Base model: Qwen/Qwen3-235B-A22B
• Minimum scale factor: 0.5
• Smoothness: 10.0
• Sample size: 1000
• Window size: 4
• N-gram weights: [0.5, 0.3, 0.2]

Unicode Ranges

• Range 1: 0x4e00 - 0x9fff
• Range 2: 0x3400 - 0x4dbf
• Range 3: 0x20000 - 0x2a6df
• Range 4: 0xf900 - 0xfaff
• Range 5: 0x2e80 - 0x2eff
• Range 6: 0x2f00 - 0x2fdf
• Range 7: 0x2ff0 - 0x2fff
• Range 8: 0x3000 - 0x303f
• Range 9: 0x31c0 - 0x31ef
• Range 10: 0x3200 - 0x32ff
• Range 11: 0x3300 - 0x33ff

Statistics

• Target tokens: 26,153
• Broken tokens: 1,457
• Modified tokens: 27,564